【作者】 石晓龙；

【导师】 许进；

【作者基本信息】 华中科技大学 ， 系统工程， 2004， 博士

【摘要】 随着DNA分子计算研究的发展,特别是对DNA计算机研究的关注,以应用于生物工程的微操作系统为基础,研究能够在微观尺度下自动进行DNA计算的生物反应芯片系统成为分子计算机研究的技术关键。其主要研究内容包括对DNA分子的操控以及对DNA分子的检测两大部分,这两个部分在DNA计算机的研究中就可以映射为如何进行操纵分子按照预定的计算目的进行生化反应以及分子计算的结果如何读出这两个问题。本文针对当前在微流控制系统、微操作系统中显微视觉伺服系统中普遍存在的轴向定位问题,结合神经网络算法和图像处理技术提出了对普通显微镜下目标物体进行三维位移标定和测量的新方法。引入神经网络算法后,对目标的轴向距离的标定过程可以通过网络的训练学习自动实现,从而使得系统能够很快适应环境条件的变化。通过标定过程,在训练结束的神经网络中实现了由目标图像的特征向量到目标轴向位置的可靠映射,与这一映射的相关知识就保存在网络的权值中,从而使系统能够依据目标的实时二维图像快速实现对目标的轴向位置测量。在介绍了DNA计算机的原理、特点及研究概况基础上,分析了微流控制系统在DNA计算机研究中的地位与作用。从DNA及蛋白质分子的操控及检测两个方面,详细介绍了微流控制系统的研究进展,介绍了微流控制系统向复杂化、空间结构三维化的发展方向。从显微镜光学系统的成像原理入手,在研究目标处于不同轴向位置的成像特征基础上,分析了显微镜光学系统对轴向测量精度的影响。最后结合图像频谱分析,研究了目标图像的频率特性与目标轴向位置的关系,说明了显微镜下目标轴向测距的原理。结合人工神经网络方法研究了对显微图像的预处理方法。提出了以BP网络为基础的图像边缘检测算法,以及基于Kohonen自组织竞争学习网络的图像分割算法,并出了一种可以同时实现对目标的边缘检测与分割识别的神经网络分类器。其理论基础是边缘检测算法与图像分割算法都是根据图像二维结构构造特征矢量对图像进行分类,因此两种算法可以使用相同的输入,并且在训练结束后具有相同的拓扑结构。所不同的是前者对输入样本灰度值的突变敏感,而后者则对样本的灰度均值敏感。运用这种分类器对显微镜下目标图像进行预处理,成功地获取了与目标三维特征密切相关的目标图像的边缘及纹理信息,避免了图像背景区域对后面将要进行的三维特征分析的干扰。 <WP=5>对显微结构下目标图像的三维特征的分析。通过傅立叶变换分析了焦点位置清晰图像、离焦位置模糊图像的频率特征,得到了图像频域图中的高频分量与图像的清晰程度正相关的结论,进而基于图像的二维傅立叶变换提取了显微镜下目标的轴向距离特征标量。应用小波分析方法对显微镜下不同轴向位置目标图像的高频细节特征的作进一步分析。通过二维图像的多分辨率小波分解从目标的不同轴向位置下获得的图像中分离出图像的高频细节部分,分析结果表明图像的高频细节部分包含了与图像清晰程度密切相关的图像边缘及纹理信息,即包含了目标的轴向位置相关特征。将香农信息熵理论引入对不同轴向位置目标图像的特征分析中,通过对不同轴向位置目标图像的一维灰度熵、二维灰度—灰度均值熵以及改进后的二维灰度—梯度熵对比分析,得到了采用改进的拉普拉斯算子的二维灰度-梯度熵特征对不同轴向位置目标图像的区分结果最优,同时不同轴向位置目标图像的熵特征在焦点位置附近随目标运动方向呈单调变化,因此具备一定的区分目标距离焦点位置方向的能力。在显微镜下目标图像的三维特征分析基础上,分别应用傅立叶变换方法、小波包分解方法从目标的二维图像中提取了其三维特征向量,并以从等间距目标采样图像中提取的特征向量数据作为样本,运用神经网络方法实现了对目标轴向距离的初步测量和误差分析。最后,结合对目标图像的灰度变化特征的分析提出了加权熵特征判据,实现了对目标轴向位置方向的判别,从而可以将采集的目标图像样本集合划分为焦点上、下两个子集,依据目标图像的频率特征分别标定其焦点位置上下的轴向距离,避免了显微镜成像光学系统非线性带来的系统误差,从而最终实现了对目标轴向位置的准确测量,并通过大量实验对采集的样本图像数量、质量对系统的稳定性和测量精度的影响作了详细分析。更多还原

【Abstract】 With the development of the research of DNA computing, especially the concentration on realization of DNA Computer, Micro-flow System, a kind of Micro-Electro-Mechanical Systems (MEMS) which is extensively applied in biotech has became the key of the realization of DNA Computer. Typically a Micro-flow System is the integration of mechanical elements, sensors, actuators, and electronics on a common silicon substrate through the utilization of microfabrication technology. Purposes of the research on Micro-flow System include the manipulation of macromolecular (such as DNA and protein) and the detection of the experimental result, which can be mapped as the operation process and the read-out process in DNA computer.This dissertation is concerned with the research on the 3-Dimension measurement of objects under microscope, which is essential to the detection and manipulation problems in Micro-flow System. The author proposed a new method to detected the 3-Dimension information under microscope. By introducing neural network theory, the calibration procedure can be implement automatically by neural network. After the calibration procedure, the relationship between micrographic images and lengthwise distance of the object under microscope is mapped as the weight point of the neural network, so the measurement procedure can be quickly implement automatically with an input image of the object under microscope.By introduceing the theory, traits and developments of DNA computer in brief, put forward the importance of Micro-flow System in the developments of DNA computer. Then by discussing investigations of Micro-flow System in manipulation and detection of DNA and protein molecular, proposes that the structure of Micro-flow System is tend to be complication and 3-Dimension. Finally summarized the latest evolves of 3-D detection under microscope and give the main result of the dissertation as well. Based on the image theory of microscope, studies the relationship between object positon and the image difinition of microscope through geomtetrical-optical approximation, and discusses the certainty of microscope measurement. Finally gives the measurement theory of microscope by analzing <WP=7>the image spectrum of microscope.Studies the application of neural network in pre-process of micrographic images. The author proposed a new edge detection network through BP algorithm, applied Kohonen SOFM on image segmentation, and finally organized these two algorithms to a neural network classifier for the pre-process of micrographic images.Work over the 3-D related feature extraction from pre-processed microscope images, analysis the relationship between frequency features of micrographic images and the position of objects under microscope, proposed methods to extract features related with lengthwise distance from 2D micrographic images.The dissertation gives a new measurement system to detection the lengthwise position of objects under microscope with conclusion. This measuerment system use a novel entropy criterion to determinate the direction of object position, thus divides the samples into two subset so that the sample characteristic value be valid in measuring object’s lengthwise position, and avoid the nonlinear abrration of microscope optical system on a certain degree. Finally the stabilization and accuracy standard of this new measurement system is checkout by a lager quantity of experimental data as well.更多还原